Wind storage combined frequency modulation control method based on model predictive control and deep learning

Abstract

The application relates to the technical field of new energy power grid coordinated control, and discloses a wind storage combined frequency modulation control method based on model predictive control and deep learning. The method comprises the following steps: driving a high-frequency updated system feature library, constructing a dynamic digital mapping of a wind storage system, and generating a control instruction according to multi-level scenario evolution and collaborative optimization. The control instruction is verified and decomposed, and then is sent to physical equipment for execution. System response data are synchronously collected, frequency domain feature extraction and pattern matching are carried out, and the dynamic digital mapping is triggered for closed-loop adaptive adjustment. The method realizes online dynamic updating of the control model and accurate correction of the system frequency domain characteristics, so that the model can closely track actual object changes, thereby improving the control precision and dynamic response quality of the wind storage combined frequency modulation.

Description

Technical Field

[0001] This invention relates to the field of coordinated control technology for new energy power grids, specifically to a wind-storage joint frequency regulation control method based on model predictive control and deep learning. Background Technology

[0002] As wind power penetration in power systems continues to increase, its output volatility and uncertainty place significant pressure on system frequency stability. Utilizing wind-storage integrated systems for frequency regulation has become an important technological direction. Current mainstream methods are mostly based on model predictive control frameworks, relying on pre-built fixed mathematical models for optimization. However, the actual operating state of power systems, equipment characteristics, and external environment are constantly changing, making it difficult for fixed model parameters to accurately describe the real-time dynamics of the integrated system, leading to discrepancies between model predictions and actual behavior. This model mismatch directly reduces the effectiveness of control commands and affects frequency regulation performance.

[0003] In existing technologies, model correction largely relies on the time-domain error feedback of the system output after the control command is executed, adjusting model parameters by minimizing the time-domain trajectory deviation. However, the core objective of frequency control is to maintain system frequency stability and possess good dynamic quality, which is mainly reflected in the system's frequency domain characteristics. Fine-tuning the model solely based on time-domain errors cannot guarantee the model's accuracy in reflecting key frequency domain characteristics such as inertial response and damping characteristics. Therefore, the generated commands may not optimally suppress frequency oscillations, limiting further improvements in control quality.

[0004] To address the two major shortcomings of models being unable to adapt to dynamic changes in the system online and model corrections not being directly targeted at the core frequency domain target of frequency modulation, a new control method is needed that can enable online model adaptation and perform targeted self-correction based on frequency domain characteristics. Summary of the Invention

[0005] The purpose of this invention is to provide a wind-storage joint frequency regulation control method based on model predictive control and deep learning, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a wind-storage joint frequency regulation control method based on model predictive control and deep learning, the method comprising:

[0007] A dynamic digital mapping for a wind-storage integrated system is constructed, and the dynamic digital mapping is driven by a frequently updated system feature library.

[0008] Based on the dynamic digital mapping, multi-level scenario evolution is performed in the time domain to generate a multi-level wind-storage evolution state set;

[0009] Based on the multi-level wind-storage evolution state set, a two-level collaborative optimization is performed within the model predictive control framework to generate an initial collaborative control instruction set;

[0010] The initial cooperative control instruction set is subjected to spatiotemporal consistency verification and decomposition to obtain the verified underlying execution instructions;

[0011] Based on the verified underlying execution instructions, control the physical wind storage equipment to perform frequency modulation actions;

[0012] The system response data after the frequency modulation action is executed is collected synchronously, frequency domain feature extraction and mode matching are performed, and then the adaptive adjustment of dynamic digital mapping is triggered.

[0013] Preferably, the dynamic digital mapping for constructing the wind-storage integrated system is driven by a frequently updated system feature library, including:

[0014] Real-time output data of wind turbine generators, multi-dimensional status data of energy storage devices, and frequency data of grid interconnection nodes are collected in parallel to form a raw data stream. The raw data stream is then synchronized with timestamp alignment and anomaly cleaning is performed to obtain clean time-series data. A deep learning network is used to perform feature mining and correlation analysis on the clean time-series data to extract core feature vectors characterizing the dynamic relationship between wind and energy storage coupling. These core feature vectors are combined with the physical topology parameters of the wind-energy storage combined system to instantiate an interactive digital system model in a virtual simulation environment. This digital system model forms the basis of the dynamic digital mapping. A feature library update mechanism is established to periodically or event-triggeredly fuse and iterate the newly extracted core feature vectors with historical features to update the system feature library, thereby driving the synchronous evolution of the dynamic digital mapping.

[0015] Preferably, the step of performing multi-level scenario evolution in the time domain based on the dynamic digital mapping to generate a multi-level wind-storage evolution state set includes:

[0016] Starting from the current state of the dynamic digital mapping, the predicted uncertainty trajectory of wind power and the dynamic response constraint trajectory of the energy storage system are extracted. Based on the predicted uncertainty trajectory and the dynamic response constraint trajectory, multiple evolution scenario series with different time granularity and spatial coverage are constructed. In each evolution scenario series, a preset grid frequency disturbance mode is injected, and forward rolling simulation is performed using the digital system model. During each rolling simulation, the state variable change history of the wind-storage joint system at all key nodes is recorded. The simulation results of all evolution scenario series are aggregated and classified, and organized according to time scale and state dimension to finally form the multi-level wind-storage evolution state set, which contains potential system state paths from the second level to the minute level.

[0017] Preferably, the step of performing two-layer collaborative optimization within the model predictive control framework based on the multi-level wind-storage evolution state set to generate an initial collaborative control instruction set includes:

[0018] The multi-level wind-storage evolution state set is analyzed and transformed into a state reference sequence in the prediction time domain within the model predictive control framework. A two-layer optimization objective is designed, with the inner optimization objective focusing on minimizing the operational losses and state deviations of the energy storage devices, and the outer optimization objective focusing on smoothing the system frequency fluctuations revealed by the multi-level wind-storage evolution state set. Within each control cycle of the model predictive control framework, the state reference sequence is compared with the current actual system state to generate a state error sequence. Using the state error sequence as input, and combining it with the two-layer optimization objective, an optimization problem with multi-time-period constraints is solved. The decision variables of the optimization problem are the charging and discharging power command sequences of each energy storage node over a future period. The solved charging and discharging power command sequences are combined and encoded with the expected wind power prediction values ​​for the corresponding time periods to generate the initial collaborative control command set.

[0019] Preferably, the step of performing spatiotemporal consistency verification and decomposition on the initial cooperative control instruction set to obtain verified underlying execution instructions includes:

[0020] A spatiotemporal consistency verification rule base is established, which includes power balance constraints, ramp rate constraints, equipment safe operating range constraints, and cross-regional coordination constraints. The initial collaborative control instruction set is projected onto a virtual spatiotemporal plane, and its compliance with all constraints in the spatiotemporal consistency verification rule base is checked. If a conflict is found, an instruction correction process is initiated, which eliminates the conflict by adjusting the time phase or spatial allocation of the instructions until all constraints are satisfied. The verified and corrected collaborative control instructions are decomposed into specific control parameters for individual wind turbine generators or energy storage converters according to the actual control interface protocol of the physical equipment. These specific control parameters constitute the verified underlying execution instructions.

[0021] Preferably, controlling the physical wind storage equipment to perform frequency modulation actions according to the verified underlying execution instructions includes:

[0022] The verified underlying execution instructions are sent to the corresponding field control units via an industrial communication network; the field control units convert the received specific control parameters into drive signals to control power electronic switching devices or prime mover adjustment mechanisms; during instruction execution, key electrical quantities, including output voltage, current and frequency, are monitored to ensure that the actual output is within the allowable deviation range from the instruction requirements; the precise start time, duration and actual output energy value of each frequency modulation action are recorded to form an action execution log.

[0023] Preferably, the step of synchronously acquiring system response data after the frequency modulation action is executed, performing frequency domain feature extraction and mode matching, and then triggering the adaptive adjustment step of dynamic digital mapping includes:

[0024] After the frequency regulation action is executed, the dynamic response curve of the power grid frequency and the power injection curve of the key nodes of the wind-storage combined system are acquired through a wide-area measurement system. The acquired dynamic response curve of the power grid frequency is transformed in the frequency domain to extract its characteristic frequency components and attenuation coefficients. The extracted characteristic frequency components and attenuation coefficients are matched with typical frequency regulation response patterns pre-stored in the pattern library. Based on the matching results, the actual frequency regulation performance of the current control loop is evaluated and a performance evaluation report is generated. The performance evaluation report is correlated with the action execution log to generate suggestions for adjusting the feature extraction weights of the deep learning network or the optimization objective function parameters in the model predictive control framework, thereby triggering the adaptive adjustment of the dynamic digital mapping.

[0025] Preferably, the adaptive adjustment step of the dynamic digital mapping specifically includes:

[0026] The adjustment suggestions are received and input as incremental learning samples into the deep learning network. Inside the deep learning network, the backpropagation algorithm is used to fine-tune the network connection weights to better capture the system dynamic features that lead to the current frequency modulation performance. At the same time, according to the parameter part in the adjustment suggestions, the cost coefficients or constraint boundaries involved in the optimization problem in the model predictive control framework are updated. After the update is completed, the dynamic digital mapping enters the next working cycle to process new real-time data with the adjusted internal parameters.

[0027] Preferably, the multi-level scenario evolution process in the time domain further includes an environmental factor coupling step:

[0028] High-precision numerical weather forecast data for the region where the wind-storage integrated system is located is acquired. The numerical weather forecast data includes spatiotemporal grid prediction information of wind speed, wind direction, air pressure, and temperature fields. The numerical weather forecast data is used as boundary conditions and coupled to the aerodynamic module of the wind turbine generator in the digital system model, thereby generating wind power output fluctuation scenarios strongly correlated with environmental factors in the evolution scenario spectrum. The wind power output fluctuation scenarios are integrated into the multi-level wind-storage evolution state set for subsequent optimization.

[0029] Preferably, the process of performing two-layer collaborative optimization within the model predictive control framework further includes an energy storage lifetime state feedback step:

[0030] The system receives health status indicators from the management systems of each energy storage device in real time. These indicators include the current cycle count, internal resistance growth rate, and capacity retention rate. The health status indicators are quantified into lifetime loss cost coefficients and dynamically injected into the inner-layer optimization objective. This allows the optimization problem with multi-time-period constraints to automatically balance frequency regulation performance requirements with long-term lifetime loss of energy storage devices, thereby reflecting a protective scheduling strategy for energy storage devices in the initial coordinated control command set.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] By establishing a dynamic digital mapping driven by a frequently updated system feature library, traditional static models with fixed parameters or offline identification are replaced. This mapping can update its internal features and mapping relationships online and dynamically based on the latest system operational response data, ensuring that the model closely follows the dynamic characteristic drift of the actual wind-storage integrated system caused by changes in equipment status, operating point migration, and external environmental disturbances. This makes predictions and scenario evolution based on this mapping more reflective of the system's real-time behavior, improving prediction accuracy and providing a reliable decision-making basis for subsequent optimized control. This, in turn, enhances the adaptability and control precision of the integrated frequency regulation system under complex and variable operating conditions.

[0033] After control is executed, system response data is synchronously acquired, and frequency domain features are extracted and pattern matched. The matching results are then used to trigger adjustments to the dynamic digital mapping. This constitutes a closed-loop adaptive correction loop oriented towards frequency modulation performance. The triggering and basis for the adjustment are directly derived from the actual dynamic characteristics of the system in the frequency domain, rather than simple time-domain trajectory tracking errors. This ensures that the self-improvement direction of the digital mapping is highly consistent with the core objective of frequency modulation control, enabling the model to more accurately characterize and predict the key dynamic behaviors of the system under frequency disturbances. Consequently, the commands generated by the controller can optimize the system's frequency response characteristics and improve damping to suppress oscillations. Attached Figure Description

[0034] Figure 1 This is a schematic diagram illustrating the working principle of the wind-storage joint frequency regulation control method based on model predictive control and deep learning described in this invention.

[0035] Figure 2 A flowchart for constructing a dynamic digital mapping of a wind-storage integrated system;

[0036] Figure 3 A flowchart for executing multi-level scenario evolution in the time domain;

[0037] Figure 4 The percentage distribution of voltage, current and frequency deviations within the wind-storage joint frequency regulation control cycle;

[0038] Figure 5 A comparison chart of frequency response indicators under the combined wind and energy storage frequency regulation control cycle. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Please see Figure 1 This invention provides a wind-storage joint frequency regulation control method based on model predictive control and deep learning. The method includes: constructing a closed-loop adaptive control system, the operation of which begins with the construction of a dynamic digital map of the wind-storage joint system. The dynamic digital map is a virtual model driven by a frequently updated system feature library, reflecting the dynamic characteristics of the physical system in real time. Based on this dynamic digital map, multi-level scenario evolution is performed in the time domain, generating a set of multi-level wind-storage evolution states covering different time scales and potential paths. Based on this state set, two-level collaborative optimization is performed within the model predictive control framework to generate an initial set of collaborative control instructions. This instruction set needs to undergo spatiotemporal consistency verification and decomposition, transforming it into verified execution instructions that can directly drive the underlying equipment. The physical wind-storage equipment executes frequency regulation actions according to these instructions. After the actions are executed, the system synchronously collects response data, performs frequency domain feature extraction and pattern matching, thereby triggering the adaptive adjustment of the dynamic digital map itself, thus forming a closed loop of continuous learning and optimization.

[0041] Example 1: See Figure 2The system collects real-time output data from wind turbine generators, multi-dimensional status data from energy storage devices, and frequency data from grid interconnection nodes in parallel. These data are aggregated to form a raw data stream. The raw data stream undergoes synchronization and timestamp alignment, and outlier data points are removed to obtain clean time-series data. Deep learning networks are used to perform feature mining and correlation analysis on the clean time-series data, extracting core feature vectors characterizing the dynamic relationship between wind and energy storage coupling. These core feature vectors are combined with the physical topology parameters of the wind-energy storage system to instantiate an interactive digital system model in a virtual simulation environment. This digital system model forms the basis of the dynamic digital mapping. A feature library update mechanism is established, periodically or event-triggeredly fusing and iterating the newly extracted core feature vectors with historical features to update the system feature library, thereby driving the synchronous evolution of the dynamic digital mapping. In some embodiments, the parallel acquisition process is accomplished through sensors deployed in the wind turbine nacelle, an energy storage battery management system, and a synchronous phasor measurement device in the power grid dispatch center. The real-time output data of the wind turbine includes rotor speed, pitch angle, and instantaneous active power. The multi-dimensional status data of the energy storage device includes state of charge, terminal voltage, internal temperature, and health status signals. The frequency data of the grid interconnection node is the system frequency deviation value sampled with millisecond-level precision. These data streams are transmitted in parallel to the central processing unit via high-speed industrial Ethernet to form a raw data stream containing multi-source heterogeneous information.

[0042] In practice, synchronizing and aligning the original data stream with timestamps and cleaning outlier data are key steps to ensure data quality. Timestamp alignment is based on international standard time protocols to synchronize data packets from different acquisition terminals with time accuracy down to the microsecond level. Outlier cleaning uses algorithms based on sliding windows and statistical outlier detection to identify and remove outlier points caused by momentary communication interruptions or sensor failures, ultimately obtaining clean time-series data that is strictly aligned on the timeline and has valid data points.

[0043] It is understandable that using deep learning networks to perform feature mining and correlation analysis on clean energy time-series data is the core of extracting the core feature vector. The deep learning network adopts a long short-term memory network structure with a multi-head attention mechanism. This network takes the wind turbine output sequence, energy storage device state sequence, and grid frequency sequence from the clean energy time-series data as parallel inputs. Through the nonlinear dynamic coupling relationship learned during the multi-layer network learning process, the network outputs a fixed-dimensional vector, namely the core feature vector. Each dimension element in the core feature vector maps a potential dynamic feature of the wind-storage integrated system under a specific operating mode. The feature extraction process can be characterized by the following formula:

[0044]

[0045] in: This represents the extracted core feature vector. This represents the feature transformation function performed by a long short-term memory network with an attention mechanism. , and These represent the cleaned data matrices for wind turbine generators, energy storage devices, and power grid frequency, respectively. This represents the set of trainable weight parameters for a deep learning network.

[0046] Optionally, the core feature vector is combined with the physical topology parameters of the wind-storage combined system and a digital system model is instantiated in a virtual simulation environment. The physical topology parameters include the electrical connection relationship between the wind turbine generator and the energy storage converter, the line impedance parameters, and the transformer ratio. In the virtual simulation environment, the core feature vector is used to parameterize a pre-built electromechanical transient model of the wind-storage combined system with a standard architecture, so that the dynamic response characteristics of the model match the behavior of the current actual physical system. The interactive digital system model instantiated in this way constitutes the basis of dynamic digital mapping. This model can receive control commands and output simulation results consistent with the expected behavior of the physical system.

[0047] In some embodiments, a feature library update mechanism is established to drive the synchronous evolution of dynamic digital mapping. The feature library is organized using a circular buffer and a hierarchical index structure. The periodic update strategy is set to push the latest extracted core feature vector into the feature library buffer every five minutes. When the system detects a severe disturbance in the power grid frequency that exceeds a threshold, an event-driven update is immediately triggered. During each update, the new core feature vector is weighted and fused with the feature vectors in the most recent historical time window in the buffer. The fusion weights are adaptively allocated according to the intensity of the system dynamics at the corresponding moment of the feature vector. The fused feature vector set is subjected to dimensionality reduction and redundancy removal through principal component analysis, and the latest version of the feature library is iteratively generated. The update of the system feature library directly drives the weight fine-tuning of the deep learning network and the refresh of the feature parameters of the digital system model in dynamic digital mapping, thereby completing the overall synchronous evolution of dynamic digital mapping.

[0048] Example 2: See Figure 3Starting from the current state of the dynamic digital mapping, the project extracts the predicted uncertainty trajectory of wind power and the dynamic response constraint trajectory of the energy storage system. Based on these trajectory, multiple evolutionary scenario series with different temporal granularities and spatial coverage are constructed. High-precision numerical weather prediction data for the region where the wind-storage integrated system is located are acquired, including spatiotemporal grid predictions of wind speed, wind direction, air pressure, and temperature fields. The numerical weather prediction data is used as boundary conditions and coupled to the aerodynamic module of the wind turbine generator in the digital system model, thereby generating wind power output fluctuation scenarios strongly correlated with environmental factors within the evolutionary scenario series. A pre-defined grid frequency disturbance mode is injected into each evolutionary scenario series, and forward rolling simulations are performed using the digital system model. During each rolling simulation, the state variable changes of the wind-storage integrated system at all key nodes are recorded. The inference results of all evolution scenario series are aggregated and classified, and organized according to time scale and state dimension to form a multi-level wind-storage evolution state set, which contains potential system state paths from the second level to the minute level. Wind power output fluctuation scenarios are integrated into this state set for subsequent optimization.

[0049] In practice, the predicted uncertainty trajectory of wind power and the dynamic response constraint trajectory of the energy storage system are extracted from the current state of the dynamic digital mapping. The predicted uncertainty trajectory of wind power is generated by a deep learning network based on historical output sequences and real-time meteorological data. It is represented as a confidence interval band centered on the predicted power value and changing over time. The dynamic response constraint trajectory of the energy storage system comes from the real-time state parameters of the energy storage device management system. It defines the feasible domain range of the energy storage device in terms of charging and discharging power, energy boundary and ramping capability in the future.

[0050] In some embodiments, multiple evolution scenario families with different time granularities and spatial coverage are constructed based on the predicted uncertainty trajectory and the dynamic response constraint trajectory. The time granularity includes the second, ten-second, and minute levels, and the spatial coverage covers a single energy storage unit, a cluster level, and the entire wind-storage combined site level. Each evolution scenario family consists of a set of simulation paths that share the same initial conditions but have different evolution parameters. The evolution parameters are sampled from the confidence interval of the predicted uncertainty trajectory and the boundary conditions of the dynamic response constraint trajectory according to a specific strategy. The construction process can be characterized by the following formula:

[0051]

[0052] in: Represents a set of evolutionary scene lineages. Represents a single evolutionary scenario. This represents a time granularity parameter that belongs to a preset time granularity set. , The set of evolutionary parameters representing the scene. From predicting uncertain trajectories and dynamic response constraint trajectory Defined probability distribution Generated by sampling from the middle.

[0053] It is understandable that high-precision numerical weather forecast data of the area where the wind-storage integrated system is located is obtained and coupled into the wind turbine aerodynamic module of the digital system model as boundary conditions. The high-precision numerical weather forecast data is provided in the form of spatiotemporal grids with a horizontal resolution of kilometers and a temporal resolution of minutes. The data includes predictions of wind speed, wind direction, air pressure and temperature fields for the next few hours. The coupling process is completed by interpolating the wind speed and wind direction data at the grid points to the hub height position of the wind turbine and inputting the air pressure and temperature data into the air density correction model of the wind turbine aerodynamic module, thereby generating wind power output fluctuation scenarios that are strongly correlated with environmental factors in the evolution scenario spectrum.

[0054] In practice, a preset grid frequency disturbance mode is injected into each evolution scenario spectrum and forward rolling simulation is performed using a digital system model. The preset grid frequency disturbance mode comes from a historical frequency event database, including various typical modes such as load step disturbance, generator disconnection disturbance, and periodic fluctuation disturbance. The simulation starts from the current dynamic digital mapping state. Within each simulation step, the digital system model solves the system differential algebraic equations based on the wind power output fluctuation scenario, energy storage dynamic response constraints, and injected grid frequency disturbance set in the evolution scenario spectrum, and forward rolling simulation of the system state.

[0055] Optionally, during each rolling simulation, the state variable change history of the wind-storage integrated system at all key nodes is recorded. Key nodes include the wind turbine generator output bus, the energy storage device access point, and the grid connection point. The recorded state variables include node voltage, frequency, active power output, reactive power output, and the state of charge of the energy storage device. Each simulation will generate a system state time series corresponding to a specific evolution scenario spectrum and a specific frequency disturbance mode.

[0056] In some embodiments, the inference results of all evolution scenario genealogies are aggregated and classified, and organized according to time scale and state dimension to finally form a multi-level wind-storage evolution state set. The aggregation operation gathers all state time series from different time granularities, different spatial coverage and different frequency disturbance modes into a unified state database. The classification operation indexes and groups the state series in the database according to time scale labels and state dimension labels. The time scale labels correspond to the second level, the ten-second level and the minute level, and the state dimension labels correspond to power, voltage, frequency and state of charge. The final multi-level wind-storage evolution state set is a structured data set that contains potential system state paths from the second level to the minute level. Wind power output fluctuation scenarios that are strongly correlated with environmental factors are integrated into the multi-level wind-storage evolution state set for subsequent optimization.

[0057] Example 3: The multi-level wind-storage evolution state set is analyzed and transformed into a state reference sequence in the prediction time domain within the model predictive control framework. A two-layer optimization objective is designed: the inner optimization objective focuses on minimizing the operational losses and state deviations of the energy storage devices, while the outer optimization objective focuses on smoothing the system frequency fluctuations revealed by the multi-level wind-storage evolution state set. Health status indicators from the management systems of each energy storage device are received in real time, including the current cycle count, internal resistance growth rate, and capacity retention rate. These health status indicators are quantified into lifetime loss cost coefficients and dynamically injected into the inner optimization objective. Within each control cycle of the model predictive control framework, the state reference sequence is compared with the current actual system state to generate a state error sequence. Using the state error sequence as input, combined with the two-layer optimization objective, an optimization problem with multi-time-period constraints is solved. The decision variable of the optimization problem is the sequence of charging and discharging power commands for each energy storage node over a future period. This allows the optimization problem to automatically balance frequency regulation performance requirements with the long-term lifetime loss of energy storage devices when solving the multi-time-period constraint optimization problem, thereby reflecting a protective scheduling strategy for energy storage devices in the initial coordinated control command set. The obtained charging and discharging power command sequence is combined and encoded with the expected value of wind power prediction for the corresponding time period to generate the initial collaborative control command set.

[0058] In practical implementation, the multi-level wind-storage evolution state set is analyzed and transformed into a state reference sequence in the prediction time domain within the model predictive control framework. The multi-level wind-storage evolution state set is a structured set containing multiple potential system state paths. The analysis process extracts all state path segments from the multi-level wind-storage evolution state set, starting from the current moment and covering a complete prediction time domain in the future, based on the timestamp of the current control cycle. These state segments are then aligned and averaged in the time dimension to generate a smooth state trajectory, i.e., the state reference sequence, which serves as an optimization reference target. Each point in the state reference sequence contains the expected values ​​of the system frequency, node power, and energy storage state of charge at the corresponding future moment.

[0059] In some embodiments, a two-layer optimization objective is designed, wherein the inner optimization objective focuses on minimizing the operating losses and state deviations of the energy storage device, and the outer optimization objective focuses on smoothing the system frequency fluctuations revealed by the multi-level wind-storage evolution state set. The mathematical expression of the inner optimization objective focuses on penalizing the frequent fluctuations in the charging and discharging power of the energy storage device and the deviation of the state of charge from the ideal operating point, while the mathematical expression of the outer optimization objective focuses on minimizing the integral sum of squares of the system frequency deviation. The two objectives are related through a coupling variable, which represents the power support required by the energy storage device to smooth frequency fluctuations.

[0060] It is understandable that the system receives real-time health status indicators from the management systems of various energy storage devices. These indicators include the current cycle count, internal resistance growth rate, and capacity retention rate. The energy storage device management system calculates these indicators in real time using internal sensors and battery models, and reports them periodically via the communication network. The internal resistance growth rate is the percentage change relative to the rated internal resistance, and the capacity retention rate is the ratio of the current available capacity to the initial rated capacity. In specific implementation, the health status indicators are quantified into lifetime loss cost coefficients and dynamically injected into the inner-layer optimization objective. The quantification process is completed through a preset mapping function. The inputs to the mapping function are the current cycle count, internal resistance growth rate, and capacity retention rate, and the output is a lifetime loss cost coefficient greater than zero. The lifetime loss cost coefficient is multiplied by the unit power operation cost of the energy storage device and added as an additional cost term to the inner-layer optimization objective function. This makes the charging and discharging behavior of the energy storage device appear "more costly" in the optimization problem when the health status of the energy storage device is worse. Therefore, when solving optimization problems with multi-time-period constraints, the optimization problem can automatically balance the frequency regulation performance requirements with the long-term lifetime loss of the energy storage device, thus reflecting a protective scheduling strategy for the energy storage device in the initial coordinated control command set.

[0061] Optionally, within each control cycle of the model predictive control framework, the state reference sequence is compared with the current actual system state to generate a state error sequence. The current actual system state consists of real-time measurement values ​​collected by the wide-area measurement system and the local monitoring system, including the actual frequency of the power grid, the actual output of the wind farm, and the actual state of charge of the energy storage. The state error sequence is a sequence consisting of the difference between the expected value at each future time point in the state reference sequence and the current actual system state in the corresponding dimension.

[0062] In some embodiments, a state error sequence is used as input, combined with a two-layer optimization objective, to solve an optimization problem with multi-time-period constraints. The decision variables of the optimization problem are the charging and discharging power command sequences of each energy storage node over a future period. The multi-time-period constraints include upper and lower power limits of the energy storage device, energy capacity constraints, charging and discharging efficiency constraints, and system power balance constraints. The solution process employs a sequential quadratic programming algorithm with interior point method. Under the condition of satisfying all multi-time-period constraints, an optimal set of charging and discharging power command sequences for the energy storage nodes is found, minimizing the overall objective function composed of the inner and outer layer objectives. The overall objective function can be characterized by the following formula:

[0063]

[0064] in: This represents the decision variable, namely the future sequence of charging and discharging power commands for each energy storage node. Indicates the decision variables The predicted frequency deviation sequence derived from the system state, This represents the outer optimization objective function, used to penalize frequency bias. , This represents the inner optimization objective function, used to penalize the energy storage power sequence. Charged state sequence and its health status cost coefficient The resulting losses and These are weighting coefficients used to coordinate inner and outer layer objectives, and health status cost coefficients. It consists of the lifespan loss cost coefficient of energy storage equipment.

[0065] In practice, the obtained charging and discharging power command sequence is combined and encoded with the wind power prediction expectation value for the corresponding time period to generate an initial collaborative control command set. The wind power prediction expectation value is obtained from numerical weather forecasting and wind power prediction model. The combined encoding packages the energy storage charging and discharging power command sequence and the wind power prediction expectation value according to timestamp and node number to form a set of structured commands containing time, location, power value and control type, i.e., the initial collaborative control command set.

[0066] Example 4: Establish a spatiotemporal consistency verification rule base, which includes power balance constraints, ramp rate constraints, equipment safe operating range constraints, and cross-regional coordination constraints. The initial collaborative control instruction set is projected onto a virtual spatiotemporal plane, and its compliance with all constraints in the spatiotemporal consistency verification rule base is checked. If conflicts are found, an instruction correction process is initiated. This process eliminates conflicts by adjusting the time phase or spatial allocation of the instructions until all constraints are satisfied. The verified and corrected collaborative control instructions are decomposed into specific control parameters for individual wind turbine generators or energy storage converters according to the actual control interface protocol of the physical equipment. These specific control parameters constitute the verified underlying execution instructions. The verified underlying execution instructions are sent to the corresponding field control units via an industrial communication network. The field control units convert the received specific control parameters into drive signals to control power electronic switching devices or prime mover regulating mechanisms. During instruction execution, key electrical quantities, including output voltage, current, and frequency, are monitored to ensure that the actual output is within the allowable deviation range from the instruction requirements. The precise start time, duration, and actual output energy value of each frequency regulation action are recorded to form an action execution log.

[0067] In practical implementation, a spatiotemporal consistency verification rule base is established. This rule base includes power balance constraints, ramp rate constraints, equipment safe operating range constraints, and cross-regional coordination constraints. The power balance constraint requires that the algebraic sum of all commanded power within the same time segment be equal to the grid dispatch command or local power plan. The ramp rate constraint restricts the rate of change of power commands between adjacent control cycles of wind turbine generators and energy storage devices from exceeding the maximum ramp capacity allowed by their physical equipment. The equipment safe operating range constraint defines the upper and lower limits of active power output, reactive power output range, and voltage operating range of each controlled device. The cross-regional coordination constraint is used to ensure that the power exchange commands between sub-clusters do not exceed the transmission capacity of the interconnection lines when the wind-storage joint system contains multiple geographically dispersed sub-clusters.

[0068] In some embodiments, an initial cooperative control instruction set is projected onto a virtual spatiotemporal plane and checked to see if it satisfies all constraints in the spatiotemporal consistency check rule base. The virtual spatiotemporal plane is a two-dimensional coordinate system, with the horizontal axis representing the time axis and the vertical axis representing the device or node axis. Each instruction in the initial cooperative control instruction set is mapped to a point or a line segment in the coordinate system according to its execution time and target device. The checking process uses a check algorithm to traverse each constraint in the spatiotemporal consistency check rule base and verify whether the relationship between all mapped points and line segments violates the mathematical expression of the constraint. For example, it checks whether the sum of the power of all device instructions is zero at a certain moment to check the power balance constraint. The check process can be characterized by the following formula:

[0069]

[0070] in: Indicates time and device nodes The set of instruction values ​​on, This represents a constraint function that performs calculations on a set of instruction values. This represents the limit value vector corresponding to the constraint.

[0071] It is understandable that if a conflict is detected, an instruction correction process is initiated. The instruction correction process eliminates the conflict by adjusting the time phase or spatial allocation of the instructions. Adjusting the time phase of the instructions involves shifting some or all of the power instructions of a certain device forward or backward by several control cycles on the time axis. Adjusting the spatial allocation of the instructions involves redistributing the power instructions that exceed the safe operating range constraints of a certain device to other devices with adjustment margins in the same time segment. The correction process is an iterative optimization process until all constraints are satisfied.

[0072] In practical implementation, the verified and corrected collaborative control commands are decomposed into specific control parameters for individual wind turbine generators or energy storage converters according to the actual control interface protocol of the physical devices. These specific control parameters constitute the verified underlying execution commands. For wind turbine generators, the specific control parameters include target active power setpoints, reactive power setpoints, or pitch angle reference values; for energy storage converters, the specific control parameters are typically active power command values, charge / discharge flags, and voltage reference values. The decomposition process is based on a preset mapping table, which defines the conversion relationship between abstract power commands in the collaborative control commands and specific device executable parameters. Refer to Table 1 for a command decomposition mapping table.

[0073] Table 1: Decomposition and Mapping Table of Cooperative Control Commands to Specific Control Parameters

[0074] Cooperative control instructions (abstract) Target device type Specific control parameters Communication Protocol Node_A,P=+1.5MW,T1 Doubly fed wind turbine generator set Active power setpoint = 1500kW, reactive power mode = constant voltage IEC61400-25 ESS_Unit_3,P=-0.8MW,T2 Lithium battery energy storage converter Power command = -800kW, control mode = constant power Modbus TCP Cluster_B,Q=+2.0MVar,T1-T5 Full-power converter fan cluster Reactive power setpoint = 2000kVar, time curve takes effect DNP3.0

[0075] Optionally, the verified underlying execution instructions are sent to the corresponding field control unit via an industrial communication network. The industrial communication network adopts a TCP / IP-based power dispatch data network or a field industrial Ethernet. The verified underlying execution instructions are encapsulated into data frame format according to the communication protocol defined in the mapping table and sent to the field control unit address of the target device through the network interface.

[0076] In some embodiments, the field control unit converts the received specific control parameters into drive signals and controls the power electronic switching devices or prime mover regulating mechanism. For energy storage converters, the conversion process involves generating the active power command value into the switching signal of the insulated gate bipolar transistor through a pulse width modulation algorithm. For wind turbine generators, the conversion process involves converting the pitch angle reference value into the actual mechanical angle of the blade through a hydraulic or electric actuator. In practical implementation, the specific implementation method of the pulse width modulation algorithm is reflected in the active power command value processing stage of the energy storage converter in the field control unit. The field control unit integrates a digital signal processor or a dedicated pulse width modulation controller, which continuously samples the received active power command value digital signal. The core of the algorithm uses the active power command value as a modulation wave and compares it with an internally generated high-frequency triangular carrier signal in real time. The comparator circuit generates a series of square wave signals with pulse widths proportional to the magnitude of the command value. These square wave signals constitute the switch drive signal, which is directly applied to the gate port of the insulated gate bipolar transistor after level conversion and isolation amplification circuits to control its on and off states. By adjusting the duty cycle of the pulse sequence, the fundamental component of the AC side output voltage of the energy storage converter is precisely controlled, thereby realizing rapid and smooth adjustment of active power output to meet the response requirements of dynamic power injection of energy storage equipment when the grid frequency fluctuates. In practice, key electrical quantities are monitored during command execution to ensure that the actual output is within the allowable deviation range from the command requirements. These key electrical quantities include output voltage, current, and frequency. Monitoring is performed through the analog input module built into the field control unit or a separate measurement unit. The monitored data is compared with the command requirements in real time. The allowable deviation range is predefined by the equipment technical specifications; for example, the tracking error of the active power command must be kept within ±1% of the rated power. Essentially, the precise start time, duration, and actual output energy value of each frequency modulation action are recorded to form an action execution log. The precise start time is taken from the timestamp of successful command issuance and confirmation by the field control unit. The duration is calculated from the precise start time until the actual output power returns to the command-required steady-state value or a new command is received. The actual output energy value is obtained by integrating the actual power curve during command execution. The action execution log is stored in a historical database for subsequent analysis.

[0077] See Figure 4 In the wind-storage combined frequency regulation control process, this figure shows the percentage changes in voltage deviation, current deviation, and frequency deviation under different control cycles (T0-T9), with 1.0% as the upper limit of allowable deviation. Specifically, the voltage, current, and frequency deviations in each control cycle are all within the range of 0.09%-0.42%, not exceeding the upper limit of allowable deviation, reflecting that the adopted wind-storage combined frequency regulation control method can effectively control the electrical quantity deviations within a reasonable range in each control cycle.

[0078] Example 5: After the frequency regulation action is executed, the dynamic response curve of the power grid frequency and the power injection curve of key nodes of the wind-storage combined system are acquired through a wide-area measurement system. The acquired dynamic response curve of the power grid frequency is transformed in the frequency domain to extract its characteristic frequency components and attenuation coefficients. The extracted characteristic frequency components and attenuation coefficients are then matched with typical frequency regulation response patterns pre-stored in a pattern library. Based on the matching results, the actual frequency regulation performance of the current control cycle is evaluated, and a performance evaluation report is generated. The performance evaluation report is correlated with the action execution log to generate suggestions for adjusting the feature extraction weights of the deep learning network or the parameters of the optimization objective function in the model predictive control framework, thereby triggering adaptive adjustments to the dynamic digital mapping. The adjustment suggestions are received and input as incremental learning samples into the deep learning network. Inside the deep learning network, the backpropagation algorithm is used to fine-tune the network connection weights to better capture the system dynamic characteristics that lead to the current frequency regulation performance. Simultaneously, based on the parameter part of the adjustment suggestions, the cost coefficients or constraint boundaries involved in the optimization problem in the model predictive control framework are updated. After the update is completed, the dynamic digital mapping enters the next working cycle to process new real-time data with the adjusted internal parameters.

[0079] In specific implementation, after the frequency regulation action is executed, the dynamic response curve of the power grid frequency and the power injection curve of the key nodes of the wind-storage combined system are acquired through a wide-area measurement system. The phasor measurement unit of the wide-area measurement system synchronously records the voltage phase and amplitude of each interconnection point of the power grid at a sampling frequency higher than 100 Hz, and calculates the accurate system frequency change curve accordingly. The power injection curve of the key nodes of the wind-storage combined system is obtained in real time by power transmitters installed at the outlet of the wind turbine generator and the grid connection point of the energy storage device. The two curves are strictly aligned on the time axis based on a unified clock source. In some embodiments, the acquired dynamic response curve of the power grid frequency is subjected to frequency domain transformation and its characteristic frequency components and attenuation coefficients are extracted. The frequency domain transformation adopts the short-time Fourier transform or wavelet transform method to decompose the frequency deviation sequence in the time domain into the energy distribution in the time-frequency domain. The characteristic frequency components refer to one or more oscillation frequencies that dominate the energy distribution. The attenuation coefficient is obtained by exponential fitting of the envelope of a specific oscillation frequency component to quantify the damping characteristics of frequency fluctuations.

[0080] It is understandable that the extracted characteristic frequency components and attenuation coefficients are matched with typical frequency modulation response patterns pre-stored in a pattern library. The pattern library stores various standard response patterns summarized from historical frequency modulation events or simulation experiments. Each standard response pattern is defined by a set of nominal values ​​of characteristic frequencies and attenuation coefficients and their allowable fluctuation range. The matching process calculates the multidimensional Euclidean distance or cosine similarity between the currently extracted characteristic frequency components and attenuation coefficients and each standard response pattern. The matching process can be characterized by the following formula:

[0081]

[0082] in: This indicates the similarity matching score. Indicates the currently extracted number Each characteristic frequency component, This indicates the first standard response pattern in the pattern library. One characteristic frequency nominal value, This parameter represents the allowable fluctuation range of the characteristic frequency component. Indicates the currently extracted number One attenuation coefficient, This indicates the first standard response pattern in the pattern library. Nominal value of attenuation coefficient This parameter represents the allowable fluctuation range of the attenuation coefficient. and These represent the number of characteristic frequency components and the attenuation coefficient, respectively.

[0083] In practice, the actual frequency regulation performance of the current control loop is evaluated based on the matching results, and a performance evaluation report is generated. The highest similarity score in the matching results and its corresponding standard response mode identify the type of the current system response, such as "underdamped oscillation", "overdamped recovery" or "ideal critical damping". The performance evaluation report is a structured document, which includes the response type determination, the calculated values ​​of key performance indicators and the description of the deviation relative to the expected target. Key performance indicators include the maximum frequency deviation, settling time and regulation deviation. Optionally, the performance evaluation report and action execution log are correlated and analyzed to generate suggestions for adjusting the feature extraction weights of the deep learning network or the parameters of the optimization objective function in the model predictive control framework. The correlation analysis uses a data fusion platform to align the response type and deviation information in the performance evaluation report with the precise instruction sequence in the action execution log through timestamp alignment and causal correlation. The analysis aims to identify the potential relationship between the control instruction pattern and the generation of non-ideal frequency modulation performance. Based on this relationship, adjustment suggestions are generated. The adjustment suggestions may include increasing the attention weights for a certain type of temporal feature in the deep learning network or modifying the coefficient of the penalty frequency deviation term in the outer optimization objective function of the model predictive control framework, thereby triggering adaptive adjustments of the dynamic digital mapping.

[0084] See Figure 5 In a wind-storage joint frequency regulation control method based on model predictive control and deep learning, the quantitative results of maximum frequency deviation (Hz), settling time (s), and regulation deviation (Hz) under different control cycles are presented. Specifically, the figure shows the distribution of two core frequency regulation performance indicators with the control cycle as the horizontal axis: the orange bars represent the maximum frequency deviation, whose value generally decreases as the control cycle progresses, reflecting the gradually increasing suppression effect of the control strategy on the frequency fluctuation amplitude; the purple bars represent the settling time, whose trend is related to the maximum frequency deviation, reflecting the time characteristic of the system recovering from frequency disturbance to a stable state; and the regulation deviation (the small data at the bottom of the figure) characterizes the steady-state accuracy of the frequency regulation action. The dynamic changes of these indicators correspond to the adaptive adjustment stage of dynamic digital mapping. Through frequency domain feature extraction and pattern matching, the optimization parameters of the model predictive control framework and the feature weights of the deep learning network are continuously fine-tuned, ultimately making the frequency regulation performance indicators under each control cycle exhibit the expected evolution law.

[0085] In some embodiments, adjustment suggestions are received and input as incremental learning samples into a deep learning network. Inside the deep learning network, the backpropagation algorithm is used to fine-tune the network connection weights to better capture the system dynamic features that lead to the current frequency modulation performance. The incremental learning process retains the knowledge gained from historical training and trains on the new adjustment suggestion samples with a small learning rate, so that the feature extraction capability of the deep learning network can adapt to the slow changes in the system's operating characteristics.

[0086] In practical implementation, the cost coefficients or constraint boundaries involved in the optimization problem in the Model Predictive Control Framework (MPC) are updated according to the parameter portion of the adjustment suggestions. If the adjustment suggestions point to the parameters of the objective function, the corresponding weight coefficients in the outer or inner objective function of the MPC are directly replaced. If the adjustment suggestions point to the constraint boundaries, the boundary value arrays of the upper and lower limits of energy storage device power or the ramp rate constraint are modified when modeling the optimization problem. The update operation is completed in the parameter configuration module of the control system. It can be understood that after the update is completed, the dynamic digital mapping enters the next working cycle and processes the new real-time data with the adjusted internal parameters. The next working cycle means starting from the parallel acquisition of real-time output data of wind turbine generators, multi-dimensional state data of energy storage devices, and frequency data of grid interconnection nodes. The new round of control loop will be based on the feature extraction weights of the fine-tuned deep learning network and the updated MPC parameters, thereby realizing the continuous adaptive evolution of the dynamic digital mapping.

[0087] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A wind-storage joint frequency regulation control method based on model predictive control and deep learning, characterized in that, This is achieved through the following steps: A dynamic digital mapping for a wind-storage integrated system is constructed, driven by a frequently updated system feature library, including: Real-time output data of wind turbine generators, multi-dimensional status data of energy storage devices, and frequency data of grid interconnection nodes are collected in parallel to form a raw data stream. The raw data stream is then synchronized with timestamp alignment and anomaly cleaning is performed to obtain clean time-series data. A deep learning network is used to perform feature mining and correlation analysis on the clean time-series data to extract core feature vectors characterizing the dynamic relationship between wind and energy storage coupling. These core feature vectors are combined with the physical topology parameters of the wind-energy storage combined system to instantiate an interactive digital system model in a virtual simulation environment. This digital system model forms the basis of the dynamic digital mapping. A system feature library update mechanism is established, periodically or event-triggered, to fuse and iterate the newly extracted core feature vectors with historical features to update the system feature library, thereby driving the synchronous evolution of the dynamic digital mapping. Based on the dynamic digital mapping, multi-level scenario evolution is performed in the time domain to generate a multi-level wind-storage evolution state set, including: Starting from the current state of the dynamic digital mapping, the predicted uncertainty trajectory of wind power and the dynamic response constraint trajectory of the energy storage system are extracted. Based on the predicted uncertainty trajectory and the dynamic response constraint trajectory, multiple evolution scenario series with different time granularity and spatial coverage are constructed. In each evolution scenario series, a preset grid frequency disturbance mode is injected, and forward rolling simulation is performed using the digital system model. During each rolling simulation, the state variable change history of the wind-storage joint system at all key nodes is recorded. The simulation results of all evolution scenario series are aggregated and classified, and organized according to time scale and state dimension to finally form the multi-level wind-storage evolution state set, which includes potential system state paths from the second level to the minute level. Based on the multi-level wind-storage evolution state set, a two-level collaborative optimization is performed within the model predictive control framework to generate an initial collaborative control instruction set; The initial cooperative control instruction set is subjected to spatiotemporal consistency verification and decomposition to obtain the verified underlying execution instructions; Based on the verified underlying execution instructions, control the physical wind storage equipment to perform frequency modulation actions; The system response data after the frequency modulation action is executed is collected synchronously, frequency domain feature extraction and mode matching are performed, and then the adaptive adjustment of dynamic digital mapping is triggered.

2. The wind-storage joint frequency regulation control method based on model predictive control and deep learning according to claim 1, characterized in that, Based on the multi-level wind-storage evolution state set, a two-level collaborative optimization is performed within the model predictive control framework to generate an initial collaborative control instruction set, including: The multi-level wind-storage evolution state set is analyzed and transformed into a state reference sequence in the prediction time domain within the model predictive control framework. A two-layer optimization objective is designed, with the inner optimization objective focusing on minimizing the operational losses and state deviations of the energy storage devices, and the outer optimization objective focusing on smoothing the system frequency fluctuations revealed by the multi-level wind-storage evolution state set. Within each control cycle of the model predictive control framework, the state reference sequence is compared with the current actual system state to generate a state error sequence. Using the state error sequence as input, and combining it with the two-layer optimization objective, an optimization problem with multi-time-period constraints is solved. The decision variables of the optimization problem are the charging and discharging power command sequences of each energy storage node over a future period. The solved charging and discharging power command sequences are combined and encoded with the expected wind power prediction values ​​for the corresponding time periods to generate the initial collaborative control command set.

3. The wind-storage joint frequency regulation control method based on model predictive control and deep learning according to claim 2, characterized in that, The process of performing spatiotemporal consistency verification and decomposition on the initial cooperative control instruction set to obtain verified underlying execution instructions includes: A spatiotemporal consistency verification rule base is established, which includes power balance constraints, ramp rate constraints, equipment safe operating range constraints, and cross-regional coordination constraints. The initial collaborative control instruction set is projected onto a virtual spatiotemporal plane, and its compliance with all constraints in the spatiotemporal consistency verification rule base is checked. If a conflict is found, an instruction correction process is initiated, which eliminates the conflict by adjusting the time phase or spatial allocation of the instructions until all constraints are satisfied. The verified and corrected collaborative control instructions are decomposed into specific control parameters for individual wind turbine generators or energy storage converters according to the actual control interface protocol of the physical equipment. These specific control parameters constitute the verified underlying execution instructions.

4. The wind-storage joint frequency regulation control method based on model predictive control and deep learning according to claim 3, characterized in that, The step of controlling the physical wind storage equipment to perform frequency modulation actions according to the verified underlying execution instructions includes: The verified underlying execution instructions are sent to the corresponding field control units via an industrial communication network; the field control units convert the received specific control parameters into drive signals to control power electronic switching devices or prime mover adjustment mechanisms; during instruction execution, key electrical quantities, including output voltage, current and frequency, are monitored to ensure that the actual output is within the allowable deviation range from the instruction requirements; the precise start time, duration and actual output energy value of each frequency modulation action are recorded to form an action execution log.

5. The wind-storage joint frequency regulation control method based on model predictive control and deep learning according to claim 4, characterized in that, The process of synchronously acquiring system response data after the frequency modulation action is executed, extracting frequency domain features and performing mode matching, and then triggering the adaptive adjustment phase of dynamic digital mapping includes: After the frequency regulation action is executed, the dynamic response curve of the power grid frequency and the power injection curve of the key nodes of the wind-storage combined system are acquired through a wide-area measurement system. The acquired dynamic response curve of the power grid frequency is transformed in the frequency domain to extract its characteristic frequency components and attenuation coefficients. The extracted characteristic frequency components and attenuation coefficients are matched with typical frequency regulation response patterns pre-stored in the pattern library. Based on the matching results, the actual frequency regulation performance of the current control loop is evaluated and a performance evaluation report is generated. The performance evaluation report is correlated with the action execution log to generate suggestions for adjusting the feature extraction weights of the deep learning network or the optimization objective function parameters in the model predictive control framework, thereby triggering the adaptive adjustment of the dynamic digital mapping.

6. The wind-storage joint frequency regulation control method based on model predictive control and deep learning according to claim 5, characterized in that, The adaptive adjustment step of the dynamic digital mapping specifically includes: The adjustment suggestions are received and input as incremental learning samples into the deep learning network. Inside the deep learning network, the backpropagation algorithm is used to fine-tune the network connection weights to better capture the system dynamic features that lead to the current frequency modulation performance. At the same time, according to the parameter part in the adjustment suggestions, the cost coefficients or constraint boundaries involved in the optimization problem in the model predictive control framework are updated. After the update is completed, the dynamic digital mapping enters the next working cycle to process new real-time data with the adjusted internal parameters.

7. The wind-storage joint frequency regulation control method based on model predictive control and deep learning according to claim 6, characterized in that, The multi-level scenario evolution in the time domain further includes environmental factor coupling steps: High-precision numerical weather forecast data for the region where the wind-storage integrated system is located is acquired. The numerical weather forecast data includes spatiotemporal grid prediction information of wind speed, wind direction, air pressure, and temperature fields. The numerical weather forecast data is used as boundary conditions and coupled to the aerodynamic module of the wind turbine generator in the digital system model, thereby generating wind power output fluctuation scenarios strongly correlated with environmental factors in the evolution scenario spectrum. The wind power output fluctuation scenarios are integrated into the multi-level wind-storage evolution state set for subsequent optimization.

8. The wind-storage joint frequency regulation control method based on model predictive control and deep learning according to claim 7, characterized in that, The process of performing two-layer collaborative optimization within the model predictive control framework also includes an energy storage lifetime state feedback step: The system receives health status indicators from the management systems of each energy storage device in real time. These health status indicators include the current number of cycles, the internal resistance growth rate, and the capacity retention rate. The health status indicators are quantified into lifetime loss cost coefficients and dynamically injected into the inner-layer optimization objectives. This allows the optimization problem with multi-time constraints to automatically balance frequency regulation performance requirements with the long-term lifespan loss of energy storage devices when solving the optimization problem, thereby reflecting a protective scheduling strategy for energy storage devices in the initial coordinated control command set.

Images